## TL;DR
When the product changes, yesterday's healthy behaviour becomes today's anomaly: a consolidated workflow cuts event counts, a new flagship feature deserves weight the old model never gave it. Patching thresholds one complaint at a time never converges. Re-derive the weights from post-change retained and churned behaviour in one pass.

```text
health score weights wrong after product change: recalibration failed
```

## Use this when
- Scores swung after a product launch, consolidation, or packaging change
- CSMs report scores contradicting what they see in accounts
- Old top signals lost volume because the product moved

## Not for this skill when
- Only one component is broken, which is a debugging task not a recalibration
- The product change is cosmetic and usage patterns are unchanged

## Steps
1. Freeze the behaviour window: define pre and post change periods and exclude the migration weeks. Verify: clean comparison windows are set
2. For post-change data, rank signals by how well they separate retained from churned accounts. Verify: an evidence-ranked signal list replaces the inherited one
3. Draft new weights from that ranking, keeping the total scale and bands familiar to CSMs. Verify: a candidate model exists on the same 0 to 100 scale
4. Backtest the candidate on post-change history and shadow it live for one scoring cycle. Verify: backtest separation improves and shadow scores look sane to account owners
5. Cut over with a changelog, an expected-movement note per segment, and a rollback snapshot. Verify: the new weights ship explained, with the old model restorable

## Variant phrasings
### health weights stale after product change
A related phrasing for the same underlying problem.
### recalibrate scoring after launch
A related phrasing for the same underlying problem.
### old usage signals lost meaning
A related phrasing for the same underlying problem.

Compatibility: Compatibility: any health scoring stack (warehouse models or CS platform rules) with retained/churned history, current versions.

## Why it happens
Weights encode what the product used to reward. A material product change invalidates that encoding at once, so incremental tweaks chase a moving target. Re-deriving from post-change outcomes is one honest rebuild instead of fifty quiet distortions.

## Edge cases / pitfalls
- Post-change history is short at first. Prefer fewer, stronger signals over a rich model fitted on thin data.
- Migration weeks poison both windows: behaviour there is neither old nor new normal. Excluding them is not optional.
- Keep bands and scale stable so CSM playbooks tied to thresholds still make sense after recalibration.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_GrC58heNV4J8RS8c7JCAyg